MLXr turns Macs into OpenAI endpoints

MLXr is less about novel model tech and more about packaging local MLX inference into something developers can actually live with day to day. That makes it a pragmatic bet on Apple Silicon becoming a serious agent-hosting platform, not just a hobbyist playground.

• –The strongest hook is compatibility: existing OpenAI SDK clients can be pointed at `http://localhost:8000/v1` instead of being rewritten for a new stack.
• –The built-in Hugging Face browser, auto context-window detection, and persisted model settings target the annoying operational gaps that usually keep local setups feeling brittle.
• –Tool-calling normalization across Qwen, Llama, DeepSeek, Mistral, Phi, and Hermes matters more than it sounds; agent workflows tend to break on exactly this kind of model-specific formatting mismatch.
• –The project is very early, with a tiny GitHub footprint so far, which means the concept is promising but production maturity is still unproven.
• –For developers already on Apple Silicon, MLXr sits in the same workflow lane as Ollama and llama.cpp servers, but with a sharper focus on MLX-native UX and OpenAI-client drop-in use.